Efficient Calibration for RRAM-based In-Memory Computing using DoRA

TL;DR

This work tackles accuracy degradation in RRAM-based In-Memory Computing caused by conductance drift by introducing a DoRA-based calibration, which shifts the learning to a small set of SRAM-stored parameters while keeping RRAM weights fixed. The method uses feature-based knowledge distillation to guide layer-wise calibration and employs Weight-Decomposed Low-Rank Adaptation (DoRA) to adjust outputs with a compact parameter set, achieving high accuracy with minimal data. Empirical results on ResNet-50/ImageNet-1K show 69.53% restoration using only 10 calibration samples and updating just 2.34% of parameters, while avoiding RRAM writes and dramatically reducing training time and energy. This approach improves calibration efficiency, extends RRAM endurance, and enables scalable edge deployments for DoRA-enabled RIMC systems.

Abstract

Resistive In-Memory Computing (RIMC) offers ultra-efficient computation for edge AI but faces accuracy degradation due to RRAM conductance drift over time. Traditional retraining methods are limited by RRAM's high energy consumption, write latency, and endurance constraints. We propose a DoRA-based calibration framework that restores accuracy by compensating influential weights with minimal calibration parameters stored in SRAM, leaving RRAM weights untouched. This eliminates in-field RRAM writes, ensuring energy-efficient, fast, and reliable calibration. Experiments on RIMC-based ResNet50 (ImageNet-1K) demonstrate 69.53% accuracy restoration using just 10 calibration samples while updating only 2.34% of parameters.

Figures (6)

• Figure 1: (a) Impact of RRAM conductance relaxation on conductance values and inference accuracy in RIMC-based systems; (b) Comparison of write operations and in-memory computing performance between RRAM and SRAM; (c) Periodic calibration process for accuracy restoration in RIMC systems; (d) Proposed RRAM-friendly calibration approach using the DoRA weight mapping strategy, eliminating the need for RRAM writes.
• Figure 2: Impact of RRAM conductance Relaxation on (a) ResNet-20 and (b) ResNet-50. Relative Drift =$\frac{\sigma}{{G_{t}}}$
• Figure 3: Overview of the proposed ultra-efficient method. (a) Feature-based calibration utilizes intermediate feature maps from a GPU-based DNN to guide the calibration of the RIMC-based DNN. This additional feature-level information reduces the risk of overfitting, enabling calibration with a tiny dataset. Calibration is performed layerwise for the RIMC. (b) Principle of DoRA: DoRA parameters are stored in digital memory, accounting for only 2.34% of total parameters. During training, only DoRA parameters are updated, while the RRAM weights remain unchanged.
• Figure 4: Comparison of calibration dataset size for feature-based calibration and backpropagation when relative drift is 20% on (a) Cifar-100 dataset with r =2 and (b) ImageNet-1K dataset with r = 4.
• Figure 5: Comparison of calibration dataset size for feature-based calibration and backpropagation when relative drift is 20% on (a) Cifar-100 dataset and (b) ImageNet-1K dataset.